update

Author: adamant-pwn

cp-algo/math/cvector.hpp

@@ -65,13 +65,13 @@ namespace cp_algo::math::fft {
             } else if(n / 4 < pre_evals) {
                 return evalp[n / 4];
             } else {
-                return polar(1., std::numbers::pi / (ftype)std::bit_floor(n) * (ftype)eval_arg(n));
+                return polar<ftype>(1., std::numbers::pi / (ftype)std::bit_floor(n) * (ftype)eval_arg(n));
             }
         }
         static constexpr std::array<point, 32> roots = []() {
             std::array<point, 32> res;
             for(size_t i = 2; i < 32; i++) {
-                res[i] = polar(1., std::numbers::pi / (1ull << (i - 2)));
+                res[i] = polar<ftype>(1., std::numbers::pi / (1ull << (i - 2)));
             }
             return res;
         }();
@@ -203,7 +203,7 @@ namespace cp_algo::math::fft {
         std::array<point, pre_evals> res = {};
         res[0] = 1;
         for(size_t n = 1; n < pre_evals; n++) {
-            res[n] = polar(1., std::numbers::pi * ftype(eval_args[n]) / ftype(4 * std::bit_floor(n)));
+            res[n] = polar<ftype>(1., std::numbers::pi * ftype(eval_args[n]) / ftype(4 * std::bit_floor(n)));
         }
         return res;
     }();

cp-algo/util/simd.hpp

@@ -17,22 +17,13 @@ namespace cp_algo {
     return a < 0 ? -a : a;
     }
 
-    // https://stackoverflow.com/a/77376595
-    // works for ints in (-2^51, 2^51)
-    static constexpr dx4 magic = dx4() + (3ULL << 51);
-    [[gnu::always_inline]] inline i64x4 lround(dx4 x) {
-        return i64x4(x + magic) - i64x4(magic);
-    }
-    [[gnu::always_inline]] inline dx4 to_double(i64x4 x) {
-        return dx4(x + i64x4(magic)) - magic;
-    }
-
     [[gnu::always_inline]] inline dx4 round(dx4 a) {
-#ifdef __AVX2__
-        return _mm256_round_pd(a, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
-#else
-        return __builtin_convertvector(lround(a), dx4);
-#endif
+        return dx4{
+            std::nearbyint(a[0]),
+            std::nearbyint(a[1]),
+            std::nearbyint(a[2]),
+            std::nearbyint(a[3])
+        };
     }
 
     [[gnu::always_inline]] inline u64x4 montgomery_reduce(u64x4 x, u64x4 mod, u64x4 imod) {

verify/poly/wildcard.test.cpp

@@ -22,7 +22,7 @@ void semicorr(auto &a, auto &b) {
 }
 
 auto is_integer(auto a) {
-    static const double eps = 1e-8;
+    static const ftype eps = 1e-8;
     return cp_algo::abs(imag(a)) < eps
         && cp_algo::abs(real(a) - cp_algo::round(real(a))) < eps;
 }
@@ -33,7 +33,7 @@ string matches(string const& A, string const& B, char wild = '*') {
     if(!init) {
         init = true;
         for(int i = 0; i < 128; i++) {
-            project[0][i] = cp_algo::polar(1., (ftype)cp_algo::random::rng());
+            project[0][i] = cp_algo::polar<ftype>(1., (ftype)cp_algo::random::rng());
             project[1][i] = conj(project[0][i]);
         }
     }